EVENT DOSSIER
The researchers proposed the Trait-Aware Policy Optimization (TAPO) framework for post-training autoregressive multi-trait essay scoring models. This method decomposes rewards along two dimensions: samples and traits, combining global scoring consistency, trait-level accuracy, format validity, and trait-dependentness; at the same time, by integrating original text and trait descriptions to enhance prompt words, it provides rich semantic information for generating trait-specific scores. Experiments show that TAPO outperforms supervised fine-tuning and scalar reward optimization baselines on multiple key models, demonstrating the effectiveness and transferability of trait-aware post-training in essay scoring.
本文提出 Trait-Aware Policy Optimization (TAPO) 框架，用于后训练自回归多特质作文评分模型。该方法沿样本与特质两个维度分解奖励，结合全局评分一致性、特质级准确性、格式有效性及特质间依赖保持；同时通过融入原文本与特质描述增强提示词，为特质特定分数生成提供丰富语义信息。实验表明，TAPO 在多个骨干模型上均优于监督微调及标量奖励优化基线，证明了特质感知后训练在作文评分中的有效性与可迁移性。