Safe Error Correction for Language Models: Frozen-Base Adjustment with Capability Preservation

Gautam Kishore

We study a practical question: can a small correction module fix errors in a frozen language model's outputs without degrading its base capabilities? We propose CRN v2, a lightweig…