Brief

A new study published on arXiv benchmarks seven foundation models on Ukrainian legal text, revealing significant variations in tokenizer fertility and zero-shot performance. The research found that models like Qwen 3 are less efficient with tokens compared to Llama-family models, and that NVIDIA's Nemotron Super 3 outperforms Mistral Large despite having fewer parameters, at a lower cost. The study also highlights that few-shot prompting can degrade performance in Ukrainian, and that models struggle with legal language from the full-scale invasion era compared to pre-war texts. AI

A new research paper reveals that base AI models, unlike their instruction-tuned counterparts, are often misclassified as human by popular AI text detectors like GPTZero and Pangram. The study proposes a method called Humanization by Iterative Paraphrasing (HIP) to fine-tune base models into paraphrasers, which can then iteratively refine generated text to evade detection. This technique, tested on Llama-3 and Qwen-3 models across various sizes, demonstrates improved detector evasion while preserving semantic meaning, suggesting current detectors may be tracking instruction-tuning artifacts rather than inherent machine-generated text qualities. AI