{"slug": "the-shared-word-evaluating-language-models-on-dutch-word-puzzles", "title": "The Shared Word: Evaluating Language Models on Dutch Word Puzzles", "summary": "A new arXiv paper (2611.22935v1) introduces The Shared Word (TSW-NL), a Dutch-language benchmark of 222 human-curated word puzzles built from Dutch Wikipedia disambiguation pages, with three independent human answers per test item. In the authors' human study scores ranged from 49.5% to 61.0%, the highest-scoring open-weight model reached only 39.3% and many others scored below 20%, while GPT-4.1 (April 2025) scored 59.2% and GPT-5.6 Sol with max reasoning scored 88.8%. The benchmark is designed to be language-specific, open-vocabulary, human-curated and deterministically scored, and the authors present it as a recipe for building similar benchmarks in other languages.", "body_md": "arXiv:2611.22935v1 Announce Type: new \nAbstract: Many LLM benchmarks for languages other than English are translated, multiple-choice, or difficult to score automatically. We introduce The Shared Word (TSW-NL), a Dutch benchmark built around four core principles: it is language-specific, open-vocabulary, human-curated, and deterministically scored. Each puzzle gives three clues that describe different senses or uses of the same Dutch word. The clues are individually ambiguous, requiring the solver to combine evidence across multiple clues. We construct the benchmark from Dutch Wikipedia disambiguation pages, followed by automatic filtering and extensive manual curation. The final dataset contains 222 instances, with three independent human answers per test item. The task is clearly challenging: scores in our human study range from 49.5% to 61.0% and the highest-scoring open-weight model we tested reaches only 39.3%. Many others even score below 20%. At the same time, closed-source models show great progress: GPT-4.1 (April 2025) scores 59.2%, whereas GPT-5.6 Sol with max reasoning scores 88.8%. We believe this task provides a useful addition to Dutch model evaluation, plus a clear recipe that can be applied to construct a similar benchmark for other languages.", "url": "https://wpnews.pro/news/the-shared-word-evaluating-language-models-on-dutch-word-puzzles", "canonical_source": "https://www.machinebrief.com/news/the-shared-word-evaluating-language-models-on-dutch-word-puz-1gy1", "published_at": "2026-11-19 05:00:00+00:00", "updated_at": "2026-11-19 05:42:32.992441+00:00", "lang": "en", "topics": ["large-language-models", "ai-research", "natural-language-processing", "machine-learning"], "entities": ["The Shared Word", "TSW-NL", "GPT-4.1", "GPT-5.6 Sol", "Dutch Wikipedia", "arXiv"], "also_reported_by": [], "takeaway": "On the 222-item TSW-NL Dutch word-puzzle benchmark, GPT-5.6 Sol with max reasoning scores 88.8% versus 59.2% for GPT-4.1, while the best open-weight model tested reaches only 39.3% and human scores range from 49.5% to 61.0%.", "recommendation": {"advice": "Build non-English evals from native, open-vocabulary, deterministically scored tasks like TSW-NL's three-clue puzzles rather than translated multiple-choice, and collect three independent human answers per item to calibrate model scores.", "playbook": "evals-testing", "playbook_url": "https://wpnews.pro/playbooks/evals-testing"}, "alternates": {"html": "https://wpnews.pro/news/the-shared-word-evaluating-language-models-on-dutch-word-puzzles", "markdown": "https://wpnews.pro/news/the-shared-word-evaluating-language-models-on-dutch-word-puzzles.md", "text": "https://wpnews.pro/news/the-shared-word-evaluating-language-models-on-dutch-word-puzzles.txt", "jsonld": "https://wpnews.pro/news/the-shared-word-evaluating-language-models-on-dutch-word-puzzles.jsonld"}}