A preregistered study compared the accuracy of two language models, Borealis 2 and Gemma 4, in providing numeric values from Norwegian statutes and regulations. The research found no significant difference between the models in their frequency of providing outdated information. While Borealis 2 was further trained from Gemma 4, the study concluded that the difference in outdated values was not statistically significant. AI
IMPACT This study indicates that current language models may struggle with factual accuracy in specialized legal datasets, suggesting a need for improved data freshness and validation in AI training.
RANK_REASON Academic paper comparing two released language models on a specific benchmark. [lever_c_demoted from research: ic=1 ai=1.0]
- Borealis 2
- Eirik Botten Nicolaysen
- Gemma 4
- google/gemma-4-26b-a4b-it
- National Library of Norway
- NbAiLab/borealis2-26b-a4b-preview
- norregel
- Norwegian
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →