Retrievable Gradients: Continual Post-Training Without Cumulative Weight Drift

2026-06-14

AI Summary

This work addresses the challenge of catastrophic forgetting and degraded generalization in continual post-training, where repeated updates to shared parameters induce cumulative weight drift. To mitigate this, the authors propose ReGrad, a novel approach that treats gradients as retrievable units of knowledge. Specifically, document-specific gradients are precomputed offline and stored in a gradient bank; during inference, only gradients relevant to the input query are retrieved to perform transient weight adaptation. A bilevel meta-learning objective further reshapes these document gradients into generalizable adaptation signals aligned with downstream tasks. This design enables reversible, drift-free, and scalable knowledge injection. Experimental results demonstrate that ReGrad significantly outperforms both continual post-training (CPT) and retrieval-augmented generation (RAG) baselines across general and domain-specific benchmarks.

Abstract

Continual post-training enables models to absorb emerging knowledge after deployment, but repeatedly updating shared parameters can accumulate weight drift, potentially causing catastrophic forgetting and degrading general capabilities. Retrieval-augmented generation avoids such parameter drift, yet often lacks the depth of parametric knowledge integration. In this paper, we propose ReGrad (Retrievable Gradients), a new paradigm that treats gradients as retrievable units of knowledge. ReGrad pre-computes document-specific gradients offline, stores them in an indexed Gradient Bank, and retrieves only query-relevant gradients at inference time for temporary weight adaptation. However, raw language-modeling gradients are optimized for token-level document reconstruction rather than for query-driven knowledge use. We therefore introduce a bi-level meta-learning objective that reshapes document-derived gradients into generalizable adaptation signals for downstream tasks. Experiments across general and domain-specific settings show that \textsc{ReGrad} outperforms CPT and RAG baselines, enabling scalable and reversible parametric knowledge injection without accumulating weight drift.

Problem

• continual post-training
• weight drift
• catastrophic forgetting
• retrieval-augmented generation
• parametric knowledge

Innovation

• Retrievable Gradients
• Continual Post-Training
• Gradient Bank
• Meta-Learning
• Weight Drift