Merging monolingual pre-trained models won't work without both representational similarity and interference mitigation

Qualitative error categorisation with judge LLMs can provide a meaningful holistic evaluation alongside quantitative results

LLMs are robust to phrasal errors, less so to natural spelling errors; but choosing the right starting prompt has more impact

LLMs fail to exploit grammar rules; parallel data is more valuable for learning to translate

LLMs do not benefit from an explicit decomposition step in translation; in fact, 2 steps of self-refinement is most effective

Relativised minimality, phrase impenetrability, and antilocality can be attributed to a single minimal search algorithm