The ZTC-Judge-9B model has been released for evaluating the reliability of answers

Developers introduced ZTC-Judge-9B — a specialized language model designed for verifying the accuracy of responses. Unlike standard LLMs, the system performs evaluation in a single forward pass, without generating additional tokens.

The model achieved a score of 30 points on the internal efficiency metric, taking around 2 hours to process a request in the framework of test trials. The architecture is focused on automated fact-checking and evaluating the quality of generation of other models in real-time.

In the course of benchmarks, ZTC-Judge-9B demonstrated results above basic models when passing professional exams. However, when tested on tasks requiring deep scientific reasoning and safety evaluation in critical situations, the model showed performance below expected standards.

The absence of token generation allows the tool to be used for high-speed content filtering without significant computational costs. This solution complements the line of specialized verifiers aimed at reducing hallucinations in AI systems by introducing an intermediate stage of data verification.