We seek a postdoctoral researcher (starting date: January 1st, 2027) to conduct the first systematic investigation of reasoning circuit structure in compact multilingual language models. This is a bilateral DFKI-Inria collaboration project funded by BMFTR (SLIMS), bridging mechanistic interpretability and multilingual NLP, addressing a critical gap: existing circuit analysis focuses on large English-centric models, leaving the question of cross-lingual and cross-scale circuit transfer largely unexplored.

You will work closely with the ALMAnaCH team at Inria Paris (Prof. Benoit Sagot, Dr. Djamé Seddah) on controlled validation using backdoored model suites as ground truth. The project includes planned bilateral research visits and joint workshops, enabling direct collaboration with leading researchers in multilingual modeling and interpretability methodology.

You will apply the full mechanistic toolkit (activation patching, sparse autoencoders, natural language autoencoders, transcoders) to understand whether reasoning circuits survive compression and translate across typologically diverse languages. Your work will directly inform the design of more capable multilingual models and advance interpretability methods for linguistically diverse systems. We value rigorous methodology and open discussion of results, regardless of whether they confirm initial hypotheses or reveal surprising findings.

Importantly, you will work as part of a fully integrated team composed of members of another BMFTR-funded project, SAgA, ensuring that your circuit-level discoveries directly feed into safe agentic system design, creating a tight feedback loop between mechanistic analysis and safety architecture.

The position is embedded in the Multilinguality and Language Technology (MLT) group under the direction of Dr. Marius Mosbach, with broader departmental support from Prof. Kristian Kersting (SAINT deparment, Darmstadt) and Prof. Verena Wolf (NMM department, Saarbrücken); and in close colaboration with the ALMAnaCH team at Inria Paris (Prof. Benoit Sagot, Dr. Djamé Seddah); the primary supervisor of the position is Dr. Simon Ostermann; the position will be situated in his team "Efficient and Explainable NLP"

Responsibilities

Apply circuit analysis methods to identify and characterize reasoning circuits in small multilingual models
Investigate circuit transfer across model scales (1B-7B) and across 20-30 typologically diverse languages
Validate next-generation interpretability tools (NLAs, transcoders) in the multilingual, small-model regime
Design and conduct controlled validation experiments using backdoored model suites with known ground truth
Publish results and release trained interpretability artifacts (SAEs, NLAs, transcoders) openly

Required Background

PhD in computer science, NLP, linguistics, or machine learning
Established expertise in mechanistic interpretability (circuit analysis, activation patching, SAEs, or equivalent)
Strong background in multilingual or low-resource NLP
Proficiency in Python and deep learning frameworks
Track record of publications in top venues (ACL, EMNLP, NeurIPS, ICML, or equivalent)

Desirable Qualifications

Experience with sparse autoencoders, transcoders, or natural language autoencoders
Prior work on cross-lingual transfer or multilingual representation analysis
Familiarity with typologically diverse language families
Interest in AI safety and alignment
Experience with evaluation infrastructure and benchmarking

Benefits

We offer competitive, market-based compensation and many other benefits (Urban Sports Club, corporate benefits, a subsidy for your Jobticket, and much more)
3-year appointment with possibility of extension
Significant computational resources (GPU access)
Embedded international collaboration with Inria Paris ALMAnaCH team, including planned bilateral research visits and joint workshop participation
Direct mentorship from leading researchers in mechanistic interpretability and multilingual NLP across institutions
Active publication culture with strong support for dissemination in top venues
A rigorous, methodologically grounded research environment with genuine space for critical discussion and scientific independence
Supportive, respectful lab culture that values both scientific excellence and human well-being
Opportunity to co-supervise PhD and Master's students
Access to open-weight multilingual models and established benchmarking infrastructure

#J-18808-Ljbffr

Deutsches Forschungszentrum für künstliche Intelligenz GmbH