DGrid AI released the latest research paper PoQ-Judge, completing the closed loop of decentralized LLM quality assessment with a multi-architecture evaluation framework

The decentralized AI infrastructure network DGrid AI today released its latest research paper "PoQ-Judge," proposing a multi-architecture quality assessment framework that does not require reference answers. This means that in real deployment environments, there are often no standard answers for comparison, yet the protocol can still reliably score the quality of model responses and allocate incentives accordingly. This is a key piece that has long been missing in DGrid's decentralized LLM inference quality assessment system.

PoQ (Proof of Quality) is a consensus mechanism independently developed by DGrid, designed to prevent model providers from deploying low-quality models, fabricating data, or hiding computational costs at the protocol level, thereby ensuring service quality and pricing transparency. The DGrid team has been continuously working on PoQ and has published four research papers to date. The newly released PoQ-Judge has trained three assessment models covering different quality and cost scenarios, achieving a correlation of up to 0.747 with human scoring on the retention test set, significantly outperforming all previous reference answer-based evaluators, while reducing assessment costs by over 72% through cascading evaluation and online weight calibration.

With the implementation of PoQ-Judge, the entire process from quality assessment → scoring → incentive allocation has completely eliminated reliance on reference answers, thus establishing a closed loop for the quality of decentralized LLM inference.

DGrid AI is a decentralized AI intelligent network dedicated to building an open, transparent, and community-driven AI infrastructure. Focusing on model invocation and application experience, DGrid has launched several core products: the AI Gateway that aggregates mainstream large models globally, the one-click deployment platform for AI agents DClaw, the anonymous model competition platform AI Arena, and the intelligent model recommendation assistant Dori, providing one-stop services for developers and users. It is reported that DGrid AI's revenue has surpassed 20 million dollars in six months.