Simon Ostermann

Research profile

Research summary

Simon Ostermann is a researcher at the German Research Centre for Artificial Intelligence (DFKI) in Germany whose profile, name-resolved from OpenAlex bibliographic data, centers on natural language processing with an emphasis on LLM interpretability, explainable-AI evaluation, and large-scale corpus auditing. The snapshot shows 83 works in the last three years, with annual output climbing from 11 works in 2024 to 34 in 2025 and 38 in 2026 — the basis of a maximal activity score of 100. Field distribution splits evenly across three sub-areas — corpus auditing and cultural evaluation resources, LLM interpretability, and explainable-AI evaluation, one third each — which yields a focus score of 0 and mirrors topic counts led by topic modeling (84) and NLP techniques (71). Five Evolve-reviewed 2026 papers — including FineWeb-CLaR, a culture-language-region annotation of 30.9B FineWeb documents, and a critique of automated simulatability — position him as last author on four. A reproducibility score of 85 (mean across three assessed papers) reflects high readiness on three catalogued papers and medium readiness on two. Collaboration is mid-range (score 58, 17 distinct coauthors), anchored by repeated co-authorship with Josef van Genabith. Basic citation counts remain small (h-index 6, i10-index 4), and the output and impact dimensions are not yet scored in this snapshot.

Research keywords

Highlights

• FineWeb-CLaR: culture, language, and region annotations for 30.9B web documents

  As last author, Ostermann contributed to culture-region annotation of 30.9B FineWeb documents supporting benchmark-aligned corpus auditing. Evolve's review credits a valuable auditing resource with candid admission of weak cultural-label validity (22% human agreement); the accompanying analyses remain exploratory.

• Showing when automated simulatability fails as an explainability metric

  A persuasive self-critique showing LLM simulators can bypass explanations by leaning on task priors and leaking anonymized labels; evidence is mostly correlational and limited to models up to 15B, but the proposed fixes are actionable. The paper carries high reproducibility readiness.

• Training-free compositional steering of multilingual and behavioral attributes

  A last-authored study claiming training-free additive steering vectors compose per attribute with its own layer and strength; composition is strong for language plus jailbreak but weak for language plus conciseness, falling below a prompt baseline. High reproducibility readiness.

• Reframing tokenization as output supervision in autoregressive LMs

  Argues that output tokenization is the true supervision signal in autoregressive language models; clean decoupling experiments and probing classifiers at toy scale support the claim, while production-scale evidence remains indirect. High reproducibility readiness.

• A mechanistic account of syntax and language separation in multilingual translation

  An activation-patching study claiming multilingual LLMs fix target syntax before language before meaning via word-order heads; the design is careful, but effects are small and lack statistical testing, yielding medium reproducibility readiness.

Timeline

1. 2013Earliest year with a recorded work in the OpenAlex counts-by-year series (1 work); recorded output stays low through 2020, peaking at 3 works in 2018.
2. 2018A small local peak with 3 recorded works, before output falls back to 2 works in 2019 and 1 in 2020.
3. 2023Recorded output resumes at 3 works after the snapshot shows no records for 2021-2022.
4. 2024Output accelerates to 11 recorded works, the start of a steep upward trend.
5. 2025Recorded output climbs to 34 works, more than triple the 2024 count.
6. 202638 recorded works so far; all five Evolve-reviewed catalog papers are dated 2026, with Ostermann as last author on four of them.

Papers in the Evolve catalog

• 2026-09-21

  FineWeb-CLaR: Culture, Language, and Region Annotations for Benchmark-Aligned Corpus Auditing

• 2026-09-08

  Limitations of Automated Simulatability: LLM Simulators Can Bypass Explanations

• 2026-09-08

  Compositional Multilingual and Behavioral Attribute Steering

• 2026-09-01

  When Tokenization is Secretly Output Supervision

• 2026-09-01

  Separating Syntax from Language: A Mechanistic Account of Translation in Multilingual LLMs

Frequent co-authors