Unveiled in April 2022, Google Research's 540-billion-parameter Pathways Language Model (PaLM) demonstrates how large-scale autoregressive language models can perform competitive multilingual translation using few-shot prompting.

Scaling Across 6,144 TPU v4 Chips with the Pathways Architecture

In April 2022, Google Research introduced PaLM (Pathways Language Model), a dense decoder-only Transformer model scaled to 540 billion parameters and trained across two Cloud TPU v4 Pods using Google's new Pathways system. While 78% of PaLM's 780-billion-token training corpus consisted of English text, the remaining 22% spanned over 100 languages—allowing researchers to evaluate how general-purpose large language models perform on WMT machine translation benchmarks compared to specialized supervised encoder-decoder NMT systems.

Few-Shot Translation into English vs. Translation out of English

PaLM's evaluation revealed a fascinating asymmetry: when translating from foreign languages *into* English (`XX -> EN`), few-shot PaLM matched or surpassed state-of-the-art supervised NMT models on several WMT language pairs, particularly excelling on older historical benchmarks and low-resource pairs like Kazakh-to-English. Conversely, when translating *out of* English into morphologically rich target languages (`EN -> DE`, `EN -> RO`), specialized supervised NMT engines maintained a clear advantage—highlighting the enduring importance of curated parallel bitexts for target-language grammatical accuracy.

Avete bisogno di traduzioni multilingue certificate, asseverazioni legali o localizzazione conforme agli standard ISO per i vostri progetti internazionali? Richiedete un preventivo gratuito personalizzato ad Alphatrad e riceverete un piano operativo dettagliato entro 24 ore.