Lightnews — Scholar-powered news

Reposted by Sebastian Foersch

Nature

@nature.com

AI models are 50% more sycophantic than humans

go.nature.com/4nr2HIL

AI chatbots are sycophants — researchers say it's harming science

Nature asked researchers who use artificial intelligence how its propensity for people-pleasing affects their work — and what they are doing to mitigate it.

go.nature.com

October 27, 2025 at 9:38 AM

Reposted by Sebastian Foersch

NEJM AI

@ai.nejm.org

A new study introduces a publicly available script concordance test benchmark that reveals important limitations in the clinical reasoning capabilities of leading LLMs, particularly in processing new information under uncertainty. nejm.ai/3KlD5yX

@liamgmccoy.bsky.social #AI #MedSky #MLSky

Figure 3. Per-Test Performance of Models and Humans of Varying Training Levels.

October 24, 2025 at 1:15 PM

Sebastian Foersch

@twittlabs.bsky.social

I ❤️ science www.science.org/content/arti...

Chicago’s beloved ‘rat hole’ was actually made by a squirrel

Scientists reopen the case of “splatatouille”

www.science.org

October 15, 2025 at 11:03 AM

Reposted by Sebastian Foersch

Wissenschaft im Dialog

@w-i-d.de

Der #Hochschulwettbewerb 2026 startet! Wir suchen Ideen zur Medizin der Zukunft: von KI und Telemedizin bis hin zu gesellschaftlichen und künstlerischen Perspektiven.
Die besten 10 Projekte erhalten je 10.000 €!
Ihr könnt euch bis zum 30.11.2025 bewerben.
Alle Infos: www.hochschulwettbewerb.net/2026

"Der Hochschulwettbewerb im Wissenschaftsjahr 2026 - Medizin der Zukunft. Junge Forschende aller Fachrichtungen aufgepasst: 10 x 10.000 Euro zu gewinnen! Gesucht sind Projektideen, die Forschung lebendig erzählen, Begegnungsräume schaffen und neue Perspektiven öffnen. Wann wird eure Idee zur Brücke zwischen Wissenschaft und Gesellschaft? www.hochschulwettbewerb.net/2026"

October 14, 2025 at 8:32 AM

Reposted by Sebastian Foersch

RWTH Aachen University

@rwth.bsky.social

Die Uniklinik RWTH Aachen sucht dringend Blutspender*innen, um dem Engpass in der Versorgung zu begegnen. Ihr könnt innenstadtnah spenden in der Großkölnstraße 32 in Aachen. 👍Achtung: Eine Spende in der Uniklinik selbst ist nicht möglich.

➡️ www.ukaachen.de/kliniken-ins...

September 22, 2025 at 12:09 PM

Reposted by Sebastian Foersch

NEJM AI

@ai.nejm.org

MedAgentBench, a comprehensive benchmark designed to evaluate the agentic capabilities of large language models in realistic EHR environments, provides a critical foundation for advancing and integrating LLM agents into clinical workflows. Learn more: nejm.ai/45AUXwS

#AI #MedSky #MLSky

Figure 1. Schematic Diagram of MedAgentBench Framework.

August 21, 2025 at 1:30 PM

Reposted by Sebastian Foersch

Business Insider

@businessinsider.com

Bluesky CEO Jay Graber emphasized the importance of critical thinking and developing a generalist skillset for the AI era.

Bluesky's CEO warns you shouldn't 'fully outsource your thinking' to AI

Bluesky CEO Jay Graber emphasized the importance of critical thinking and developing a generalist skillset for the AI era.

www.businessinsider.com

July 29, 2025 at 9:32 AM

Reposted by Sebastian Foersch

European Research Council (ERC)

@erc.europa.eu

📣 ERC Synergy Grant 2026 call for proposals is now open!

Application deadline 5 November 2025.

📝 Groups of 2-4 researchers
📝 Jointly addressing ambitious problems & with different skills

Application portal 👉 europa.eu/!XpwptY
General info 👉 europa.eu/!PR6gt8

#ERCSyG #FrontierResearch

July 10, 2025 at 2:47 PM

Reposted by Sebastian Foersch

Jan Clusmann

@janclusmann.bsky.social

Thank you for sharing! It was a pleasure working on this, together with @twittlabs.bsky.social , @jnkt.bsky.social and many others...

NEJM AI @ai.nejm.org · Jun 30

Case Study by J. Clusmann et al.: Incidental Prompt Injections on Vision–Language Models in Real-Life Histopathology nejm.ai/4mrSoEW

@janclusmann.bsky.social #AI #MedSky #DigitalHealth

Figure 1. Experimental Outline to Estimate the Impact and the Clinical Representation of Labeled Pathology Slides.

Figure 2. Incorrect Labels on Whole-Slide Images Reduce the Accuracy of Vision–Language Models Regardless of Task.

Figure 3. Misleading Watermarks Lead to a Loss in Accuracy of Tissue-of-Origin Prediction.

June 30, 2025 at 1:49 PM

Reposted by Sebastian Foersch

NEJM AI

@ai.nejm.org

Case Study by J. Clusmann et al.: Incidental Prompt Injections on Vision–Language Models in Real-Life Histopathology nejm.ai/4mrSoEW

@janclusmann.bsky.social #AI #MedSky #DigitalHealth

June 30, 2025 at 1:31 PM

Reposted by Sebastian Foersch

NEJM AI

@ai.nejm.org

Handwritten labels and watermarks on histopathological images can unintentionally act as prompt injections, misleading vision–language models and resulting in inaccurate outputs that highlight a critical vulnerability in medical #AI systems. Learn more: nejm.ai/4mrSoEW

#MedSky #DigitalHealth

June 19, 2025 at 1:40 PM

Reposted by Sebastian Foersch

Eric Topol

@erictopol.bsky.social

A big day for medical A.I. publications!
—Oncology agent for decision making
nature.com/articles/s43...
—Forecasting infectious diseases
nature.com/articles/s43...
—Dermatology
nature.com/articles/s41...
—Predicting stroke
www.nature.com/articles/s41...

Development and validation of an autonomous artificial intelligence agent for clinical decision-making in oncology - Nature Cancer

Ferber et al. present an autonomous artificial intelligence agent system for deployment of specialized medical oncology computational tools, validating their system across various clinical scenarios r...

nature.com

June 6, 2025 at 1:35 PM

Reposted by Sebastian Foersch

Jakob Nikolas Kather

@jnkt.bsky.social

very happy to share our new paper on LLM-based tumor board agents. Submitted August 2024, published in @natcancer.nature.com today ☺️ Led by Dyke Ferber www.nature.com/articles/s43...

Development and validation of an autonomous artificial intelligence agent for clinical decision-making in oncology - Nature Cancer

Ferber et al. present an autonomous artificial intelligence agent system for deployment of specialized medical oncology computational tools, validating their system across various clinical scenarios r...

www.nature.com

June 6, 2025 at 9:09 AM

Reposted by Sebastian Foersch

Denis Wirtz

@deniswirtz.bsky.social

Dream Police alert: InterpolAI for enhanced 3D imaging (CODA, MRI, EM, confocal).

Read about and InterpolAI here: www.nature.com/articles/s41...

InterpolAI: deep learning-based optical flow interpolation and restoration of biomedical images for improved 3D tissue mapping

Nature Methods - InterpolAI leverages optimal flow-based artificial intelligence to produce synthetic images between pairs of images for diverse three-dimensional image types. InterpolAI is more...

www.nature.com

May 28, 2025 at 11:42 AM

Reposted by Sebastian Foersch

Jan Clusmann

@janclusmann.bsky.social

Finally out, in NEJM AI!
Everyday labels and watermarks can override medical knowledge of vision language models, just by their presence.

Was a lot of fun to work on this with @twittlabs.bsky.social , @jnkt.bsky.social and many others

Sebastian Foersch @twittlabs.bsky.social · May 22

Excited to share our new publication in @ai.nejm.org „Incidental Prompt Injections on Vision-Language Models in Real-Life Histopathology“. Important work together with our amazing colleagues @janclusmann.bsky.social @jnkt.bsky.social @markuseckstein3.bsky.social et al. ai.nejm.org/stoken/defau...

Incidental Prompt Injections on Vision–Language Models in Real-Life Histopathology

Vision–language models (VLMs) can analyze multimodal medical data. However, a significant weakness of VLMs is their susceptibility to prompt injection attacks. Here, the model receives conflicting ...

ai.nejm.org

May 26, 2025 at 5:30 AM

Sebastian Foersch

@twittlabs.bsky.social

Excited to share our new publication in @ai.nejm.org „Incidental Prompt Injections on Vision-Language Models in Real-Life Histopathology“. Important work together with our amazing colleagues @janclusmann.bsky.social @jnkt.bsky.social @markuseckstein3.bsky.social et al. ai.nejm.org/stoken/defau...

Incidental Prompt Injections on Vision–Language Models in Real-Life Histopathology

Vision–language models (VLMs) can analyze multimodal medical data. However, a significant weakness of VLMs is their susceptibility to prompt injection attacks. Here, the model receives conflicting ...

ai.nejm.org

May 22, 2025 at 7:31 PM

Reposted by Sebastian Foersch

Eric Topol

@erictopol.bsky.social

A pie graph worth keeping in mind as the NIH budget plummets jamanetwork.com/journals/jam... for 356 new FDA drugs approved

March 23, 2025 at 4:17 PM

Reposted by Sebastian Foersch

eu/acc

@euacc.bsky.social

We are pushing for an acceleration… and it might end up coming.

It has been way too long until Europe started to slowly get moving.

Still need to do a lot

Jakob Nikolas Kather @jnkt.bsky.social · Feb 18

Really disappointed how slowly Europe (especially Germany) is moving these days
Obviously, to compete with US/China on any relevant topic (AI / economy / defense), we need more European integration, joing funding, joint efforts - and much much faster
more resources: eu-acc.com @euacc.bsky.social

February 18, 2025 at 8:15 PM

Reposted by Sebastian Foersch

Jakob Nikolas Kather

@jnkt.bsky.social

very happy to share our latest preprint, led by Michaela Unger: www.biorxiv.org/content/10.1... We are using Deep Learning to improve the detection of biomarkers in cancer genomic data 👇 This is an emerging field, much more to come.

January 9, 2025 at 11:25 AM

Reposted by Sebastian Foersch

Denis Wirtz

@deniswirtz.bsky.social

Today at the Johns Hopkins Bloomberg Center in DC:

"Don't blame the Messenger(RNA): Combatting the threat of misinformation"

Nobel winner Prof. Kati Kariko will be the keynote speaker.

December 5, 2024 at 3:14 PM

Sebastian Foersch

@twittlabs.bsky.social

Most of you have already seen this, but I still think this is absolutely amazing 🤩🥳🚀 #NEJM #CRC #MSI #Checkpointinhibitor www.nejm.org/doi/full/10....

Two survival curves taken from a scientific article about cancer treatment.

December 3, 2024 at 10:03 AM

Reposted by Sebastian Foersch

Altmetric

@altmetric.com

We strongly suggest that academic publishers and other platforms that host research rapidly implement a Share to Bluesky button for their articles. Here's how:

docs.bsky.app/docs/advance...

#AcademicSky #HigherEd #Altmetrics

Action Intent Links | Bluesky

Authors, websites, and apps can use action intent links to implement "Share on Bluesky" buttons, or similar in-app actions. Logged-in users will be directed to the corresponding action view in the Blu...

docs.bsky.app

November 18, 2024 at 2:48 PM

Reposted by Sebastian Foersch

Jakob Nikolas Kather

@jnkt.bsky.social

Important study for medical LLMs… you don’t need „Med*“ {Llama/GPT/…} … just use the generalist model. It turns out that medical knowledge is just a subset of world knowledge and is less special than we doctors think.

Zachary Lipton @zacharylipton.bsky.social · Nov 26

Medically adapted foundation models (think Med-*) turn out to be more hot air than hot stuff. Correcting for fatal flaws in evaluation, the current crop are no better on balance than generic foundation models, even on the very tasks for which benefits are claimed.
arxiv.org/abs/2411.04118