Reproduce “Tracing Multilingual Knowledge Acquisition Dynamics in Domain Adaptation: A Case Study of English-Japanese Biomedical Adaptation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xin Zhao, Naoki Yoshinaga, Yuma Tsuta, Akiko Aizawa
This study introduces AdaXEval, an adaptive pipeline for evaluating multilingual domain knowledge acquisition in Large Language Models (LLMs) by aligning evaluation data directly with training corpora. The research details the dynamics of intralingual knowledge acquisition, including a "loss shielding" mechanism, and identifies that effective cross-lingual transfer in specialized domains critically depends on domain-specific translation signals.
Key insights
- LLMs exhibit a "loss shielding" mechanism during continual training, where correct factual sequences are protected from rapid loss growth despite overall model overfitting, ensuring continued knowledge memorization.
- Knowledge acquisition robustness is highly sensitive to training data perturbations, with lexicon-preserving changes (e.g., syntax rewriting) causing less harm than those introducing irrelevant tokens or severe structural alterations.
- Effective cross-lingual knowledge transfer in specialized domains is primarily driven by domain-specific cross-lingual signals (e.g., medical translations) rather than general-purpose multilingual corpora, yet overall gains remain limited.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.