ruben härle.

MACHINE LEARNING RESEARCHER

Ruben Härle.

PhD candidate at TU Darmstadt’s AIML Lab

I study how language models represent concepts and reason — and how we can understand and guide what happens inside them.

My work spans LLM training and pretraining data, mechanistic interpretability, and reasoning in latent spaces.

GitHubGoogle ScholarLinkedInLab profile
TRAINING & UNDERSTANDING LLMS01
Training language models and understanding their representationsConceptual illustration: pretraining data feeds a large language model. Its internal activations are examined through sparse features to reveal interpretable concepts.LLMLEARNED WEIGHTSPRETRAINING DATATRAININSPECT HIDDEN STATESACTIVATIONSSPARSEFEATURESCONCEPTSBuild the model. Understand what it learns.
From pretraining data to interpretable representations.Conceptual illustration
01 / RESEARCH

Opening the black box.

I study the internal representations of large language models, with a focus on making their behavior easier to understand and guide.

01

Interpretable representations

Sparse autoencoders, monosemanticity, and finding meaningful concepts in model activations.

02

Reasoning in latent spaces

Exploring logical reasoning within the activation spaces of language models.

03

Foundation models

Time-series representations and high-quality German-language pretraining data.

02 / SELECTED WORK

Publications & preprints

Ideas, experiments, and open research.

2026
ICLR 2026 · Time Series WorkshopBest Paper Award

Finding the Zeitgeist in Time Series Foundation Models

F. Divo, M. Kraus, R. Härle, P. Ahrend, K. Kersting

Foundation models
2024
NeurIPS 2024 · SoLaR Workshop

SCAR: Sparse Conditioned Autoencoders for Concept Detection and Steering in LLMs

R. Härle, F. Friedrich, M. Brack, B. Deiseroth, P. Schramowski, K. Kersting

Interpretability
2026
ICLR 2026 · Logical Reasoning Workshop

LLMs Gaming Verifiers: RLVR can Lead to Reward Hacking

L. Helff, Q. Delfosse, D. Steinmann, R. Härle, H. Shindo, P. Schramowski, W. Stammer, K. Kersting, F. Friedrich

Reasoning
03 / A LITTLE BACKGROUND

Research, with a
practical perspective.

I’m a PhD candidate at the AIML Lab at TU Darmstadt, where I’ve been researching machine learning since May 2024. I also spent six months as a research intern at Aleph Alpha, working on sparse autoencoders for Llama-family models.

Before my PhD, I combined my studies with industry experience at Hewlett Packard Enterprise, working across analytics, strategy, and research and development.

Outside research, wrestling has been a big part of my life. I competed in the 1st and 2nd Bundesliga and now contribute to ASV Mainz 88 through IT administration and its managing committee.

More at my university profile ↗
Ruben Härle
Ruben Härle Darmstadt, Germany

Education

M.Sc. Computer Science
TU Darmstadt · 2021–2024
Minor in discrete optimization

B.Sc. Business Informatics – Data Science
DHBW Mannheim · 2018–2021

Teaching

Deep Learning: Architectures & Methods
Teaching assistant · Summer 2026

NLP4Web
Tutor · Winters 2022/23 & 2023/24

LET’S COMPARE NOTES

Interested in understanding
what’s happening inside?