Abstrak

The reliance on single Large Language Models (LLMs) for automated academic assessment risks creating an algorithmic monoculture, where inherent model biases are amplified at scale. This paper introduces a novel framework, the Multi-Agent Consensus System (MACS), designed to mitigate this risk by simulating cognitive diversity. MACS orchestrates a heterogeneous ensemble of LLMs in a structured, adversarial peer-review workflow. The system comprises: (1) a VLM-driven multimodal extraction module for high-fidelity data retrieval from PDFs; (2) an initial review by a primary agent; (3) a critical challenge stage by secondary agents with diverse architectures; and (4) a final arbitration stage where a concluding agent synthesizes conflicting evaluations to form a robust consensus. By formalizing this process of structured disagreement and resolution, our framework moves beyond simple ensemble averaging. We introduce the Disagreement-Resolution Ratio (DRR) as a novel metric to quantify the system's ability to identify and correct initial scoring biases. Our experiments show that MACS achieves 92.3% scoring accuracy and reduces single-model scoring variance by 63%, demonstrating its superior robustness and fairness in automated academic evaluation.