Large Language Models are Parallel Multilingual Learners
Large Language Models are Parallel Multilingual Learners [50.1] 本研究では,多言語大言語モデル(LLM)の文脈内学習能力を明らかにする。 入力を複数の言語に翻訳することで、並列入力(PIM)をLLMに提供し、その理解能力を大幅に向上させる。 論文参考訳（メタデータ） (Thu, 14 Mar 2024 03:33:46 GMT)
「Considering knowledge learnt from different languages memorized in separate neurons of LLMs, a straightforward explanation for the superiority of PIM is that it leads to the increasing number of activated neurons, utilizing more knowledge during the inference stage.」はなるほどと思いつつ「This finding is similar to the synaptic pruning happening in brains, which prunes less-used neural connections and makes frequently-used neural pathways more powerful and efficient (Huttenlocher et al , 1979; Huttenlocher, 1990).」はほんまかいなと思わなくもない。