Get Started
Research questionWhen does weight-space merging preserve translation quality across models with shared versus different target languages?Independently fine-tuned translation models may use overlapping computational units while developing incompatible target-generation representations. This makes it difficult to combine their weights without losing the quality of the individual models.
AI
Machine Learning
Natural Language Processing
Latest papersRecent research connected to this question, newest first.One Model to Translate Them All? A Journey to Mount Doom for Multilingual Model MergingThe study examines fully fine-tuned language models merged for multilingual translation using shared-source, shared-target, and bidirectional consolidation settings. It evaluates representative merging strategies and analyzes internal representations, with evidence covering large-scale bilingual training and the reported language-configuration differences.research paper · Sep 3, 2026
Related questions
How can multilingual representation sharing be measured without confusing anisotropy with genuine cross-lingual structure?How can models adapt to low-resource languages without damaging source-language and related-task performance?When can simple mappings translate representations across independently trained text embedding models?How can sparse autoencoder features be shared across language models without per-model retraining?
Home
Topics
Search
Library