LLMs Get Smarter from Targeted Synthetic Multilingual Data
HOTFIXR targets the languages and prompts where a model is weakest, then generates synthetic training data to patch those gaps.
The paper frames language-specific competency as a model giving different answers to the same query depending on the prompt language. Its proposed framework probes a student model for multilingual weaknesses and builds training data around those failures. In the reported evaluations, HOTFIXR improved in-distribution performance by 6.2% on average, reduced fine-tuning-related forgetting on out-of-distribution tasks by 3.7%, and improved out-of-distribution language results by 7.1%. The authors say they will release code upon acceptance. HF Daily Papers' note
The paper frames language-specific competency as a model giving different answers to the same query depending on the prompt language. Its proposed framework probes a student model for multilingual weaknesses and builds training data around those failures. In the reported evaluations, HOTFIXR improved in-distribution performance by 6.2% on average, reduced fine-tuning-related forgetting on out-of-distribution tasks by 3.7%, and improved out-of-distribution language results by 7.1%. The authors say they will release code upon acceptance. HF Daily Papers' note
score 5