Retrievable Gradients: Continual Post-Training Without Cumulative Weight Drift

The paper introduces ReGrad (Retrievable Gradients), a novel approach for continual post-training that mitigates weight drift by storing pre-computed, document-specific gradients in an indexed Gradient Bank. At inference time, ReGrad retrieves relevant gradients for temporary weight adaptation using a bi-level meta-learning objective to optimize these gradients for generalizable task performance. This method demonstrates superior performance over existing models like CPT and RAG, enabling scalable knowledge integration while preventing catastrophic forgetting, which is crucial for practitioners working with LLMs in dynamic environments.