Native language identification is the task of inferring a non-native speaker's first language from their second-language writing. A new preprint on arXiv investigates a token-probability approach, in which each text sample is scored as a continuation of some context using the output probabilities of a large language model.

The method is described in the abstract as a way to apply LLM token probabilities to NLI, but the abstract is truncated before details of the experimental setup, datasets, or results are given. The authors have not yet provided enough information in the source to assess how the approach compares with existing NLI techniques.

As a result, the paper's contribution is clear in direction—using LLM probabilities as a signal for L1 background—but the available text leaves open questions about accuracy, robustness, and how the scoring is calibrated across different L1 groups.