LG AI Research· EXAONEResearch paper

LG AI Research presents a zero-human vision inspection framework that uses Monte Carlo Dropout for uncertainty extraction and a Vision-Language Model (VLM) for auto-labeling, achieving 93.13% accuracy with a +4.84%p improvement over random sampling. The VLM is trained through a three-stage pipeline including base model training, domain-specific fine-tuning, and reasoning reinforcement learning, and a noise-robust learning policy based on embedding space similarity is employed to handle noisy labels.

From the source

Applying this precise uncertainty sampling mechanism yielded an accuracy of 93.13%, representing a significant +4.84%p improvement compared to random sampling (88.29%) under identical data computational conditions.

lgresearch.ai

Announced

02 Sep 2026