MATE: A Multi-Agent Trait Evaluation System for Automated Essay Scoring

Automated essay scoring (AES) aims to evaluate written essays efficiently and consistently. In this context, multi-trait scoring has become particularly important because it produces fine-grained assessments of individual writing traits rather than a single undifferentiated score. Whereas earlier approaches largely rely on task-specific supervision, zero-shot solutions have emerged as a promising alternative for low-resource or rapidly changing assessment settings. However, training-free multi-trait scoring remains difficult because the system must generate trait-level judgments that are both locally well justified and globally consistent, despite strong dependencies among traits and the absence of task-specific calibration. Existing approaches often overlook these dependencies, limiting both accuracy and interpretability. To address this issue, we propose MATE, a Multi-Agent Trait Evaluation system for training-free multi-trait essay scoring. MATE employs several dimension-specific agents whose roles are defined by clustering correlated traits from scoring rubrics, thereby reducing redundancy and improving specialization. These agents independently evaluate predefined dimensions by generating rationales and scores. A Meta-cognitive Coordinator Agent (MCA) then orchestrates iterative conflict detection and discussion, resolves disagreements among dimension-specific agents, and synthesizes the final scores. By explicitly modeling interactions and dependencies among dimensions through agent-based argumentation, MATE improves both scoring coherence and explainability. Experimental results on the ASAP++ dataset show that MATE consistently outperforms strong training-free AES baselines, achieving an average QWK of 0.532 across all backbones and improving on the strongest baseline by about 5.91%. These results support the effectiveness of collaborative reasoning for capturing nuanced writing quality in zero-shot multi-trait scoring.