Prompt Engineering: Reasoning through Perspective Transition (RPT) скачать в хорошем качестве

Prompt Engineering: Reasoning through Perspective Transition (RPT)

Buy my book about prompt engineering: US: https://www.amazon.com/dp/B0GLQBS28S IN: https://www.amazon.in/dp/B0GLQBS28S UK: https://www.amazon.co.uk/dp/B0GLQBS28S BR: https://www.amazon.com.br/dp/B0GLQBS28S This research explores cross-lingual reasoning generalization in Large Reasoning Models (LRMs) to determine if mathematical skills learned in English transfer to other languages. The authors introduce the Multilingual Transferability Index (MTI) and find that while Reinforcement Post-Training (RPT) enhances reasoning, models often over-rely on English-specific patterns, hindering universal transfer. Through interventional studies, they identify a First-Parallel Leap, where adding just one parallel language during training significantly boosts multilingual performance. Furthermore, they establish a Parallel Scaling Law, proving that cross-lingual transfer follows a predictable power-law relationship with the number of training languages. Despite these gains, a Monolingual Generalization Gap remains, suggesting current models do not yet mirror the language-agnostic nature of human cognition. Ultimately, the study advocates for parallel training as a critical strategy for developing more language-agnostic and globally effective AI models.