STaR-DRO: Stateful Tsallis Reweighting for Group-Robust Structured Prediction
Researchers have developed STaR-DRO, a novel framework for improving structured prediction in large language models, particularly for tasks with imbalanced data and varying group difficulties. The framework combines modular prompt engineering with a reweighting technique that upweights persistently difficult groups without penalizing easier ones. Evaluations on the EPPC Miner task demonstrated significant improvements in zero-shot extraction and robustness when applied to Llama models, outperforming standard fine-tuning and traditional DRO methods. AI