iiinova — where your digital self keeps evolving

Homepage - Catherine Chen

Congratulations, Dr. Chen, on this achievement, and all the best for this exciting next chapter! 🎓 Follow Catherine's work here: 🔗 Personal Website: catherineschen.github.io 🔗 Google Scholar: scholar.google.com/citations?us... #FeatureFriday #PhDJourney #InformationRetrieval #ExplainableAI

catherineschen.github.io

I am a researcher who recently defended my PhD at Brown University, where I worked with Carsten Eickhoff in the Health NLP Lab. Stay tuned for where I'm headed next!

000

We are excited to see Catherine take the next step in her academic career as a postdoctoral researcher at Leiden University, where she will continue her research on how explanations affect trust in human-AI collaborative systems.

100

Her work explored what makes information relevant and how to make retrieval systems more transparent and interpretable, contributing to more effective, understandable, and accountable AI.

100

On June 15, 2026, Catherine successfully defended her PhD thesis, "Investigating Mechanisms of Relevance for Explainable Information Retrieval Systems", completing her doctoral research at Brown University.

100

Health NLP Lab @health-nlp.com · 17/08/2026

With our entire US PhD cohort now graduated, today we are celebrating Dr. Catherine Chen and her successful PhD journey! 🎉

100

If you missed #ICLR2026, here is a list of highlighted papers, chosen and recommended by Dr. Ali Bahrainian: 🔗 health-nlp.com/posts/iclr26

health-nlp.com

Health NLP | ICLR 2026, Rio de Janeiro, Brazil

000

Health NLP Lab @health-nlp.com · 06/08/2026

#ThrowbackThursday to #ACL2026 where Siran Li presented "MATCHA: Matching Text via Contrastive Semantic Alignment". Haven't read it yet? Here is a direct link to the paper: aclanthology.org/2026.finding...

000

Understanding Wacky Weights: A Dissection of SPLADE's Learned Term Importance

𝐔𝐧𝐝𝐞𝐫𝐬𝐭𝐚𝐧𝐝𝐢𝐧𝐠 𝐖𝐚𝐜𝐤𝐲 𝐖𝐞𝐢𝐠𝐡𝐭𝐬: 𝐀 𝐃𝐢𝐬𝐬𝐞𝐜𝐭𝐢𝐨𝐧 𝐨𝐟 𝐒𝐏𝐋𝐀𝐃𝐄'𝐬 𝐋𝐞𝐚𝐫𝐧𝐞𝐝 𝐓𝐞𝐫𝐦 𝐈𝐦𝐩𝐨𝐫𝐭𝐚𝐧𝐜𝐞 ✒️Gregory Polyakov, Harrisen Scells, Carsten Eickhoff arxiv.org/abs/2605.19628

arxiv.org

Learned sparse retrieval models such as SPLADE combine the effectiveness of neural architectures with the efficiency of inverted indices. As these models assign weights to terms from a fixed…

041

𝐇𝐮𝐦𝐚𝐧𝐬, 𝐋𝐋𝐌𝐬, 𝐚𝐧𝐝 𝐌𝐞𝐚𝐬𝐮𝐫𝐞𝐬 𝐃𝐨 𝐍𝐨𝐭 𝐀𝐥𝐢𝐠𝐧 𝐢𝐧 𝐀𝐭𝐭𝐫𝐢𝐛𝐮𝐭𝐞𝐝 𝐈𝐧𝐟𝐨𝐫𝐦𝐚𝐭𝐢𝐨𝐧 𝐑𝐞𝐭𝐫𝐢𝐞𝐯𝐚𝐥 ✒️Lukas Gienapp, Jenny Lang, Martin Potthast, Harrisen Scells dl.acm.org/doi/10.1145/...

121

Here are the #SIGIR2026 papers featuring our lab members: 𝐓𝐨𝐩𝐢𝐜-𝐒𝐩𝐞𝐜𝐢𝐟𝐢𝐜 𝐂𝐥𝐚𝐬𝐬𝐢𝐟𝐢𝐞𝐫𝐬 𝐚𝐫𝐞 𝐁𝐞𝐭𝐭𝐞𝐫 𝐑𝐞𝐥𝐞𝐯𝐚𝐧𝐜𝐞 𝐉𝐮𝐝𝐠𝐞𝐬 𝐭𝐡𝐚𝐧 𝐏𝐫𝐨𝐦𝐩𝐭𝐞𝐝 𝐋𝐋𝐌𝐬 ✒️Lukas Gienapp, Martin Potthast, Andrew Yates, Harrisen Scells, Eugene Yang arxiv.org/abs/2510.04633

120

Melbourne is the place to be this week! 🇦🇺 #SIGIR2026 has officially kicked off, and while we miss Carsten Eickhoff, Harrisen Scells, and Gregory Polyakov back in the office, the Health NLP Lab is well represented in Melbourne this week.

131

Tassallah Amina Abdullahi

Explore her research: 🌐 tassabdul.github.io 📚 scholar.google.com/citations?us... Congratulations, Dr. Abdullahi, on this outstanding achievement. We are excited to see where your journey takes you next and the impact your work will continue to have on trustworthy AI and healthcare. 🚀

tassabdul.github.io

PhD student at Brown University

000

Tassallah has established a strong record of impactful publications and has been an active contributor to the AI research community. Her work sits at the intersection of machine learning, natural language processing, and clinical AI, bridging technical innovation with real-world healthcare needs.

100

During her PhD, she explored questions about the reliability, reasoning, and controllability of AI systems in clinical environments. Her research contributed new insights into knowledge-grounded models, dependable inference, and mechanisms for steering AI behavior.

100

Her dissertation tackles one of the most important challenges facing modern AI: how do we ensure that systems deployed in healthcare are not only capable, but also trustworthy?

110

#FeatureFriday Today, we are celebrating a major achievement for Dr. Tassallah Amina Abdullahi! 🎉 On April 20, 2026, Tassallah successfully defended her PhD thesis, “Towards Trustworthy Clinical AI: Knowledge Grounding, Inference Reliability, and Behavioral Control.”

100

As the field continues to push the boundaries of generative AI, improving how we evaluate model outputs is just as important as improving the models themselves. We hope MATCHA contributes toward more reliable, trustworthy, and semantically grounded evaluation of AI systems.

020

We evaluated MATCHA across eight public benchmarks spanning question answering, summarization, image captioning, natural language inference, and semantic textual similarity, consistently outperforming widely used evaluation approaches and demonstrating stronger alignment with human judgments.

100

In this work, we introduce MATCHA, a novel evaluation metric that not only rewards agreement with a reference text but also explicitly penalizes contradictions. The key idea is that a good evaluation metric should recognize both what is correct and what is clearly wrong.

100

To address this gap, we present our #ACL2026 paper "MATCHA: Matching Text via Contrastive Semantic Alignment": arxiv.org/pdf/2605.27345

110

Yet many of today's most widely used evaluation metrics, such as ROUGE and BERTScore, cannot be relied on, since they can assign remarkably similar scores to statements that directly contradict one another.

100

How can we reliably evaluate whether AI-generated text is actually correct? This question is becoming increasingly important as large language models are deployed in high-stakes domains, from healthcare and education to scientific research.

100

I am currently a PhD student in Computer Science at Brown University. I am fortunate to be advised by Professor Carsten Eickhoff and Professor Ellie Pavlick. Before my PhD, I received my master’s…

We are also excited that she will be joining Microsoft as an Applied Scientist for the next stage of her career. Follow her work on her personal website: buff.ly/BKia9Ca Congratulations again to Dr. Zhang, and we wish her all the best in this exciting next chapter! 🚀 #FeatureFriday #PhDJourney

ruochenzhang.com

Ruochen Zhang

000

Through her work, Ruochen advanced research on multilingual NLP, cross-lingual transfer, and language representation in modern AI systems. She also built an impressive publication record and made impactful contributions to the NLP and machine learning community through insightful research.

100

On April 15, 2026, Ruochen successfully defended her PhD thesis, "Understanding Multilingualism in Large Language Models", culminating years of impactful research on how large language models represent, transfer, and generalize knowledge across languages and linguistic settings.

100

One by one, our PhD students at Brown University are reaching graduation milestones, and today’s post is dedicated to Dr. Ruochen Zhang! 🎉

100

When Silence Is Golden: Can LLMs Learn to Abstain in Temporal QA and Beyond?

Beyond achieving strong empirical results, the work offers valuable insights into how abstention and reasoning can be jointly optimized, paving the way toward more trustworthy and reliable LLMs. Don't miss the read: arxiv.org/abs/2602.04755

arxiv.org

Large language models (LLMs) rarely admit uncertainty, often producing fluent but misleading answers, rather than abstaining (i.e., refusing to answer). This weakness is even evident in temporal…

000

The study demonstrates that teaching LLMs when not to answer can substantially improve both reasoning performance and reliability in temporal question answering tasks.

100

Adopting a novel approach, the paper frames abstention as a teachable reasoning skill and introduces a training pipeline that combines Chain-of-Thought supervision with Reinforcement Learning guided by abstention-aware rewards.

100

This work tackles one of the most pressing challenges in LLM reliability: models often generate confident yet misleading answers instead of admitting uncertainty or abstaining when sufficient evidence is unavailable.

100

Health NLP Lab @health-nlp.com · 24/04/2026

#ThrowbackThursday to #ICLR2026, where Dr. Ali Bahreinian presented the paper "When Silence Is Golden: Can LLMs Learn to Abstain in Temporal QA and Beyond?" during the poster session.

100

The piece highlights a striking finding: many AI-generated health responses are inaccurate or incomplete, despite sounding highly credible. It’s a timely reminder of both the potential and the risks of relying on AI in healthcare. Read it here: theconversation.com/half-of-ai-h...

theconversation.com

Half of AI health answers are wrong even though they sound convincing – new study

AI chatbots can sound authoritative on health, but new research shows they often mislead, especially when users must interpret and apply the answers themselves.

010

Health NLP Lab @health-nlp.com · 24/04/2026

Need a thought-provoking read for the weekend? Don’t miss this article by Prof. Carsten Eickhoff: “Half of AI health answers are wrong even though they sound convincing,” published in The Conversation EU.

100

Reposted by Health NLP Lab

ML for Science @ml4science.bsky.social · 24/04/2026

Six days to go! Apply by April 30 for the Director of our new AI Methods & Software Hub. This is a unique opportunity to build and lead a central hub at the intersection of cutting-edge ML research and scientific applications. More information: uni-tuebingen.de/en/128980#c2...

033

Health NLP Lab @health-nlp.com · 27/03/2026

Great conference ahead! We are traveling to the Netherlands to attend #ECIR2026, and we are particularly excited about our tutorial on Mechanistic Interpretability tutorial on Sunday and the Collab-a-thons. Looking forward to insightful exchanges and constructive conversations.

010

Health NLP Lab @health-nlp.com · 13/03/2026

New video presentation available on YouTube! You can now watch Gregory Polyakov, the first author of "Interpretability Analysis of Arithmetic In-Context Learning in Large Language Models", explain their research question, methodology, and results. You can watch the video here:

youtu.be

Interpretability Analysis of Arithmetic In-Context Learning in Large Language Models

Large language models (LLMs) exhibit sophisticated behavior, notably solving arithmetic with only a few in-context examples (ICEs). Yet the computations that connect those examples to the answer…

001

Health NLP Lab @health-nlp.com · 05/03/2026

New video alert! The video presentation of the paper "A Survey on LLM-Assisted Clinical Trial Recruitment" by Dr. Shrestha Ghosh is now live on YouTube. 🎞️ Check out the video here: www.youtube.com/watch?v=fE_3... And don't forget to subscribe to our YouTube channel for more videos!

000

With this milestone achieved, she will continue her research as a postdoctoral researcher starting in March 2026. You can explore her body of work here: dblp.org/pid/322/7058... Join us in celebrating Dr. Michal Golovanevsky for her persistence, creativity, and contributions to the field of CS👏

000

Michal has a rich portfolio of publications across top venues, and her most recent paper, "Mechanisms of Prompt-Induced Hallucination in Vision-Language Models", currently under review at ACL 2026, continues this trajectory of high-impact research.

100

In her PhD years, she focused on understanding the internal workings of VLMs, with particular attention to scalability, interpretability, and control. Her work blends rigorous theory with practical insights, pushing forward how attention mechanisms are designed and understood in multimodal contexts.

100

Health NLP Lab @health-nlp.com · 18/02/2026

Heartfelt congratulations to Dr. Michal Golovanevsky! 🎉 On January 30, 2026, our PhD student at Brown University successfully defended her thesis, "Advancing Attention Mechanisms in Multimodal Deep Learning Models", marking the culmination of years of research excellence and intellectual growth.

101

Kay highlighted what it really means to translate research into clinical practice, outlined common regulatory pathways, and shared typical mistakes teams make when compliance is treated as an afterthought. The Q&A that followed showed how relevant these challenges are to many projects.

000

Health NLP Lab @health-nlp.com · 18/02/2026

Thank you to everyone who joined our invited talk last Friday. We were very happy to welcome Kay Brosien from x-cardiac GmbH and to learn from his practical insights on AI regulation in medicine.

Prof. Carsten Eickhoff introducing Kay Brosien from x-cardiac GmbH.

110

Health NLP Lab @health-nlp.com · 12/02/2026

Reminder for the talk tomorrow: see you at the Hörsaal of Maria-von-Linden-Straße 6 at 11:00 A.M.

000

✒️ Gregory Polyakov, Catherine Chen, Carsten Eickhoff 📃 dl.acm.org/doi/10.1145/... </> github.com/polgrisha/be...

000

In “Towards Best Practices of Axiomatic Activation Patching in Information Retrieval”, we analyze these pitfalls and propose concrete best practices to make activation patching a more reliable diagnostic tool for neural rankers, paving the way toward more interpretable and trustworthy IR models.

100

However, applying activation patching in IR is far from straightforward: dataset construction choices, term rareness, small score differences, and other experimental factors can strongly bias the results and lead to misleading conclusions.

100