Get Started
Home
Topics
Search
Library
Research questionHow can models adapt to low-resource languages without damaging source-language and related-task performance?Adapting a model to a low-resource language can recruit unnecessary capacity and alter behavior learned for the source language or related tasks. The central difficulty is improving target-language performance while limiting these unintended changes.
AI
Alignment & Safety
LLM Pretraining & Post-training
Machine Learning
Mechanistic Interpretability
Natural Language Processing
Research Paper
Latest papersRecent research connected to this question, newest first.CroCo: Cross-Lingual Contrastive Preference Tuning on Self-GenerationsThe study evaluates self-generation-based preference tuning with an English-trained reward model on two multilingual language models across 14 high- and low-resource languages and structured and open-ended tasks. It examines monolingual and multilingual pairing, on-policy versus off-policy responses, and offline versus online optimization; the reported evidence is limited to these models, languages, and evaluations.research paper · Sep 2, 2026Beyond Transfer Accuracy: Mechanism-Guided Controlled Adaptation for Low-Resource LanguagesThe source studies Transformer adaptation for cross-lingual sentiment transfer on NusaX and extends the analysis to XNLI, using mechanism-guided circuit discovery and targeted updates to task-relevant components. Results are reported across two model families, with evidence focused on target accuracy, source-language retention, related-task performance, and intervention behavior.research paper · Sep 2, 2026
Related questions
How can spoken language detection adapt to underrepresented accents under low-resource constraints without overfitting?How can language models improve accessibility-focused text simplification for low-resource languages when cross-lingual transfer is unreliable?How can vision-language models maintain visual recognition when modalities are missing and source training data is unavailable?How can language models personalize for many users without costly per-user model storage?