STaR-DRO: Stateful Tsallis Reweighting for Group-Robust Structured Prediction

The article introduces STaR-DRO, a novel framework for structured prediction that enhances label accuracy and robustness in large language models, particularly in high-stakes clinical tasks. It incorporates a modular prompt-engineering architecture and a new reweighting method that selectively upweights persistently hard groups without negatively affecting easier ones, leading to significant improvements in zero-shot extraction performance across Llama models, with average label F1 gains of +14.46 and further enhancements through supervised fine-tuning. This advancement is critical for practitioners focused on reliable automated communication mining in healthcare, as it addresses common pitfalls in model performance under label imbalance and varying group difficulties.