LexNeo-Bench (Zero-Shot): leaderboard
LexNeo-Bench (Hosseini-Kivanani, 2026) tests knowledge of Luxembourgish lexical borrowing: for a highlighted token in RTL news text the model labels it native, a French loan or a German loan (an English-loan label is offered as a distractor), on 3,000 balanced items from the LuxBorrow corpus. Zero-shot prompt without external context. Accuracy in percent at temperature 0, with unparseable outputs dropped; higher is better.
Metric: Accuracy (%; borrowing-type classification of a highlighted token in a Luxembourgish RTL news sentence among native, French loan and German loan (English loan offered as a distractor label); 1,000 items per class; outputs that cannot be mapped to a label are dropped; temperature 0; zero-shot prompt with a Luxembourgish-linguist persona). Source: arxiv.org. Saturation forecast: Estimated already saturated. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 3.3 70B Instruct | 34.7 | #1353 |
| 2 | Gemma 3 27B (IT) | 31.4 | #1306 |
| 3 | Gemma 3 12B (IT) | 24.5 | #1849 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=lexneo-bench-zero-shot · How It Works · Data refreshed daily, snapshot 2026-11-12.