David Thulke

Research profile

Research summary

David Thulke is a researcher at RWTH Aachen University, Germany (ORCID 0000-0002-9808-7073), working in natural language processing with adjacent activity in speech technology. His OpenAlex topic profile is dominated by language technologies: Topic Modeling (23 records), general NLP techniques (22), speech and dialogue systems (15), speech recognition and synthesis (10), and AI in service interactions (6), with the field distribution pointing at machine translation. Recorded output starts in 2017, peaks at 12 works in 2021, and settles into five works per year in 2024 and 2025 — 13 works across the recent three-year window; the snapshot lists an h-index of 4 and i10-index of 2, consistent with an early-career record. Two 2026 arXiv papers reviewed in the Evolve catalog probe how LLMs behave under fine-tuning: one compares 11 catastrophic-forgetting mitigation methods for translation, finding EWC preserves general benchmarks but not MT-specific instruction following; the other proposes knowledge-aligned supervised fine-tuning with Recall Rewrite, whose factuality gains Evolve flags as partly inflated by refusal-driven FActScore effects. He is a middle author in both, within a 10-person recent co-author network. Reproducibility evidence covers one paper at medium readiness, so this reads as an early-stage, translation- and fine-tuning-focused profile.

Research keywords

Highlights

• 11 forgetting-mitigation methods tested for LLM translation fine-tuning

  A 2026 arXiv study compares 11 catastrophic-forgetting mitigation methods for machine translation fine-tuning; Elastic Weight Consolidation preserves general benchmarks but does not preserve MT-specific instruction following. Evolve rates the single-run evidence solid and reproducibility readiness medium.

• Knowledge-aligned supervised fine-tuning for factuality

  Proposes a knowledge-aligned SFT framing with a Recall Rewrite method; factuality gains on WildHalu and Bios benchmarks are real but partly reflect refusal-driven FActScore inflation, according to Evolve's review. Reproducibility readiness is rated medium.

• From scattered early output to steady production

  OpenAlex per-year counts rise from one work each in 2017–2019 to 12 in 2021, then settle at five works in both 2024 and 2025, with 13 works across the recent three-year window — an active, ongoing research program.

Timeline

1. 2017First works appear in the OpenAlex record (1 work).
2. 2021Publication peak: 12 works recorded in a single year.
3. 2023Output dips to 2 works, the lowest since the 2017–2019 stretch.
4. 2024Activity recovers to 5 recorded works.
5. 2025Steady pace continues with 5 recorded works.
6. 2026Two reviewed arXiv papers on LLM fine-tuning for translation and factuality (middle authorship, 3 works recorded so far).

Papers in the Evolve catalog

• 2026-09-23

  Fine-Tuning LLMs for Translation: General Forgetting Mitigation Does Not Preserve MT-Specific Instruction Following

• 2026-08-31

  Stick to What You Know: A Study of Knowledge-Aligned Supervised Fine-Tuning

Frequent co-authors