Yves Scherrer

Research profile

Research summary

Yves Scherrer is a researcher at the University of Oslo working at the intersection of machine translation, computational linguistics, and large language models. The Evolve catalog currently contains two reviewed 2026 arXiv papers, both with Scherrer in the last author position — a pattern consistent with a senior or supervisory role. The first, ReMova, fine-tunes LLMs (using LoRA and NLLB-based baselines) for English→Belarusian translation and shows through matched ablations that Belarusian-specific parallel data cleaning is a key bottleneck for low-resource MT quality, although the reported gains lack significance testing. The second, TransClean, is described by Evolve as the first benchmark for detecting and extracting clean translations from noisy LLM outputs, supported by thorough analysis and high reproduction readiness, though its LLM-annotated silver labels raise circularity concerns. Together the two papers define a coherent agenda: treating translation data quality and noise as the central lever for improving MT in both low-resource settings and LLM-centric pipelines. Bibliometrically, the snapshot records 4 distinct coauthors (Mikita Pilinka, Aliaksandr Kliujeŭ, David Samuel, Shenbin Qian), a field distribution split evenly between NLP and computational-linguistics variants of machine translation, and a mean reproducibility score of 78 across the two reviewed papers. Output, impact, and activity metrics are not yet available, so this profile should be read as an early-stage, catalog-based picture rather than a full career assessment.

Research keywords

Highlights

• First benchmark for extracting clean translations from noisy LLM outputs

  Evolve's review identifies TransClean as the first benchmark for detecting and extracting clean translations from noisy LLM outputs, built around quality estimation with COMET-KIWI and backed by thorough analysis; the main caveat is that LLM-annotated silver labels raise circularity concerns. Reproduction readiness is rated high.

• Data cleaning as the bottleneck in English→Belarusian LLM translation

  ReMova fine-tunes LLMs (with LoRA and NLLB baselines) for English→Belarusian translation and adds Belarusian-specific cleaning to the fine-tuning pipeline; a matched ablation supports data quality as the key bottleneck for low-resource MT, though the reported gains lack significance tests. Reproduction readiness is rated medium.

• A research agenda centred on translation data quality

  Across both reviewed 2026 papers, the same lever recurs: noisy or under-cleaned parallel data. ReMova targets Belarusian-specific cleaning for low-resource gains, while TransClean operationalizes noise detection at benchmark scale — a coherent data-quality-first programme for LLM-era machine translation.

Timeline

1. 2026ReMova preprint appears on arXiv: LLM fine-tuning with LoRA and NLLB baselines for English→Belarusian translation, incorporating Belarusian-specific parallel data cleaning.
2. 2026TransClean preprint appears on arXiv: a benchmark for detecting and extracting clean translations from noisy LLM outputs, using quality estimation (COMET-KIWI).
3. 2026Both catalogued outputs land in machine-translation fields with Scherrer as last author among 4 recorded distinct coauthors, marking a compact senior-supervised collaboration cluster.

Papers in the Evolve catalog

• 2026-09-14

  ReMova: Fine-tuning LLMs for English to Belarusian translation

• 2026-09-10

  TransClean: A Benchmark for Detecting and Extracting Clean Translations from Large Language Model Outputs

Frequent co-authors