Tanja Baeumel

Research profile

Research summary

Tanja Baeumel is an early-career researcher in natural language processing whose work centres on the interpretability of large language models. The bibliometric record (OpenAlex) shows a sharply focused topic profile: topic modeling (30 works), general NLP techniques (24), and explainable AI (17), with smaller clusters in domain adaptation and few-shot learning (7) and adversarial robustness (4). Output has accelerated markedly — from 4 works in 2023 and 1 in 2024 to 9 in 2025 and 17 in 2026, totalling 27 works over the last three years — while citation-based metrics (h-index 1, i10-index 0) remain minimal, consistent with an early-stage profile. The Evolve-reviewed catalog captures three 2026 arXiv papers spanning this agenda. As first author of "When Tokenization is Secretly Output Supervision", Baeumel argues that output tokenization acts as the true supervision signal for autoregressive models, supported by toy-scale decoupling and probing experiments. Two middle-author collaborations probe multilingual LLMs: one on training-free compositional attribute steering (strong for language-plus-jailbreak, weaker for language-plus-conciseness), and a mechanistic account of translation in which target syntax is fixed before language and meaning via word-order heads. Collaboration is compact — 5 distinct recent coauthors, including Josef van Genabith, Simon Ostermann, Daniil Gurgurov, Mikhail Sonkin, and Hyun Gu Kang — suggesting a tight, lab-centred working pattern.

Research keywords

Highlights

• Tokenization as the hidden supervision signal of autoregressive LLMs

  First-author study arguing that output tokenization, not the surface objective, is the true supervision signal in autoregressive language models; toy-scale decoupling and probing-classifier experiments support the claim, though production-scale evidence remains indirect (Evolve verdict). Reproducibility rated high.

• Training-free compositional steering of multilingual LLM attributes

  Middle-author work showing that additive steering vectors can be composed training-free with per-attribute layer and strength settings; combinations work strongly for language-plus-jailbreak, but language-plus-conciseness falls below a prompt baseline (Evolve verdict). Reproducibility rated high.

• A mechanistic account of syntax-first translation in multilingual LLMs

  Middle-author mechanistic study using activation patching to argue that multilingual LLM translation fixes target syntax before language before meaning, via word-order attention heads; the design is careful, but effects are small and lack statistical testing (Evolve verdict). Reproducibility rated medium.

Timeline

1. 2023First year covered by this snapshot's OpenAlex per-year counts: 4 indexed works.
2. 2024Lowest output in the covered window: 1 indexed work.
3. 2025Output rebounds to 9 indexed works.
4. 2026Most productive year: 17 indexed works, including all three Evolve-reviewed arXiv papers — a first-authored study on tokenization as output supervision, plus middle-author work on compositional attribute steering and mechanistic translation analysis.

Papers in the Evolve catalog

• 2026-09-08

  Compositional Multilingual and Behavioral Attribute Steering

• 2026-09-01first author

  When Tokenization is Secretly Output Supervision

• 2026-09-01

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

Frequent co-authors