Spectral Unforgetting: Post-Hoc Recovery of Damaged Capabilities Without Retraining

TL;DR

This paper introduces DG-Hard, a spectral repair method that post-hoc recovers damaged capabilities in language models caused by fine-tuning without retraining, by filtering the weight updates using spectral techniques.

Contribution

The paper presents DG-Hard, a novel spectral filtering approach for post-hoc repair of fine-tuned models, effectively restoring capabilities while preserving beneficial improvements.

Findings

DG-Hard achieves the strongest balanced repair across multiple benchmarks.

It restores safety alignment degraded by benign fine-tuning.

Spectral residue in weight updates is a key factor in capability loss.

Abstract

Fine-tuning a language model for a target task routinely degrades capabilities the training data never explicitly threatened. We study this phenomenon, known as catastrophic forgetting, and propose a post-hoc repair solution that uses only the pretrained checkpoint $W_{\mathrm{base}}$ and its fine-tuned descendant $W_{\mathrm{ft}}$. The goal is not merely to revert the model toward the base checkpoint, but to recover capabilities damaged by fine-tuning while preserving both the target-task gains and any beneficial held-out improvements. We introduce DG-Hard, a checkpoint-only spectral repair method for the fine-tuning update $\Delta = W_{\mathrm{ft}} - W_{\mathrm{base}}$. DG-Hard treats $\Delta$ as a low-rank task-aligned signal embedded in an IID-like noise residual that gradient descent has no incentive to remove, and applies the Donoho-Gavish hard singular-value threshold to each weight-delta matrix, keeping the structured high-energy part of the update and removing the spectral bulk. This reduces repair to a closed-form SVD filtering step requiring no data-dependent tuning. A central difficulty is evaluation: average accuracy hides per-benchmark failures, while naive recovery scores reward models that simply revert toward the base. We therefore introduce a partition-conditional metric that separately tracks healing, preservation, non-damage, and target-task retention. Across $14$ (model, task) settings and nine cross-domain held-out benchmarks, DG-Hard achieves the strongest balanced repair among post-hoc baselines. DG-Hard also restores safety alignment degraded by benign fine-tuning on three independent safety axes, despite using no alignment data. These results suggest that part of fine-tuning-induced capability loss is not an unavoidable consequence of specialization, but a removable spectral residue in the weight update itself.