Fine-tuning large language models with in-context examples, a technique called IC-Train, allows translation systems to adapt to new parallel data when deployed. But according to a new arXiv preprint, it remains unclear whether these models genuinely learn the translations of rare terms from those examples. The paper sets out to formalize this problem, rather than taking successful adaptation for granted.
The authors propose a training method that uses ciphers to strengthen in-context terminology learning, particularly for translations from rare languages. The abstract suggests that this cipher-based approach could make models more reliable when encountering uncommon vocabulary in few-shot settings.
Because only the abstract is available, the specific experimental results and comparisons are not yet reported. The paper's contribution at this stage is to define the open question and offer a novel training strategy aimed at closing the gap between apparent adaptation and true term acquisition.