<ul><li>Performance of large language models depends on the composition of their training data.</li><li>Existing approaches for selecting data mixtures for LLMs can be expensive and suboptimal.</li><li>AutoMixAlign (AMA) is a new algorithm that adaptively mixes datasets during training to balance performance across tasks.</li><li>AMA outperforms standard alignment approaches and model merging methods in multitask alignment setups.</li></ul>

AutoMixAlign: Adaptive Data Mixing for Multi-Task Preference Optimization in LLMs

Discover more