By combining steering and prompting methods, it is possible to strike a balance between controllability and generation quality.
#ConferencePaper #MachineLearning #HealthNLP

This paper evaluates steering vectors in a setting that had not previously been addressed: free-form summarization. While steering vectors can effectively control the summary's properties, higher steering strength can compromise text quality.

📄 Paper: arxiv.org/abs/2506.15301

Looking forward to insightful discussions at AACL 2025!

#AACL2025 #ClinicalTrial #LLM #NLProc #ResearchHighlight

The authors provide a comprehensive analysis of trial–patient matching and examine how LLMs can support cohort discovery. The paper reviews current benchmarks, LLM-assisted methods, and evaluation practices, and highlights key challenges and promising future research directions for clinical NLP.

Clinical trial recruitment remains a major bottleneck in medical research. Trials and patient records are largely described in natural language, and there is significant potential to leverage large language models. However, existing matching approaches are often trial-specific and limited in scope.

Outside of research, he enjoys playing the clarinet 🎵, backpacking 🏕, and exploring modern art and architecture 🎨.

#MeetTheLab #PhDLife #AIResearch #HealthNLP #TeamHighlight

Gregory's work focuses on #interpretability of language models, with a particular interest in in-context learning, retrieval, and retrieval-augmented generation (#RAG). Gregory aims to uncover how these models operate internally to make them more efficient and safer.

TopoLearn opens the door to more principled, interpretable, and reliable use of machine learning in chemistry and drug discovery.

📃 A highly recommended read for anyone exploring machine learning in chemistry or drug discovery: link.springer.com/article/10.1...

In the paper “The topology of molecular representations and its influence on machine learning performance,” Florian Rottach addresses this question and develops TopoLearn, a model that predicts the effectiveness of representations based on the topological characteristics of their feature space.

However, it is still unclear what the best molecular representation is and how much the topology of a molecular representation influences the performance of a machine learning model.

⛰️ Outside research, Florian enjoys gravel biking, ice-bathing, mountain hikes, calisthenics, and traveling to sunny places.

🔬 Research Interests
His PhD sits at the intersection of Molecular Machine Learning and Topological Data Analysis, aiming to improve models and embedding spaces for in-silico drug discovery.

💼 Career Pathway
Florian holds a Master’s in Industrial Engineering from KIT and explored software, polymer tech, and automotive industries — always in ML-focused roles.
At Boehringer Ingelheim, he works with data ranging from images to molecules.

📄 Paper: health-nlp.com/files/pubs/m...

Code: github.com/leontramonti...

#ML4H2025 #AIInHealthcare #HealthNLP #ClinicalTrials #ResearchHighlight