Teaching a document parser to find its own blind spots (thesfstack.substack.com)

The SoMark team has announced a breakthrough in document parsing with their new model, OXR, which utilizes a pioneering recursive self-improvement (RSI) loop during post-training. This method allows the model to learn from its own mistakes automatically, identifying errors and uncertainties in its outputs by parsing unlabeled documents multiple times. The model's ability to analyze discrepancies and uncertainties leads to targeted training that focuses on common weaknesses rather than individual errors, ultimately enhancing its accuracy and adaptability. This innovation is significant for the AI/ML community as it addresses the common challenges in document parsing where traditional models struggle to generalize across varied document layouts. By integrating a vision-language model in the training process, OXR not only identifies errors but also generates synthetic data to fill in training gaps without manual intervention. The model has already outperformed its peers in various benchmarks, ranking first in multiple parsing competitions. The release of OXR's code and weights on GitHub encourages further experimentation and collaboration, paving the way for improved document parsing solutions across diverse applications.