dorsal/arxiv
View SchemaSafeguarding LLM Fine-tuning via Push-Pull Distributional Alignment
| Authors | Haozhong Wang, Zhuo Li, Yibo Yang, He Zhao, Hongyuan Zha, Dandan Guo |
|---|---|
| Categories | |
| ArXiv ID | 2601.07200vv1 |
| URL | https://arxiv.org/abs/2601.07200 |
| License | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ |
Abstract
The inherent safety alignment of Large Language Models (LLMs) is prone to erosion during fine-tuning, even when using seemingly innocuous datasets. While existing defenses attempt to mitigate this via data selection, they typically rely on heuristic, instance-level assessments that neglect the global geometry of the data distribution and fail to explicitly repel harmful patterns. To address this, we introduce Safety Optimal Transport (SOT), a novel framework that reframes safe fine-tuning from an instance-level filtering challenge to a distribution-level alignment task grounded in Optimal Transport (OT). At its core is a dual-reference ``push-pull'' weight-learning mechanism: SOT optimizes sample importance by actively pulling the downstream distribution towards a trusted safe anchor while simultaneously pushing it away from a general harmful reference. This establishes a robust geometric safety boundary that effectively purifies the training data. Extensive experiments across diverse model families and domains demonstrate that SOT significantly enhances model safety while maintaining competitive downstream performance, achieving a superior safety-utility trade-off compared to baselines.
{
"annotation_id": "fe471084-c9de-4f5a-8235-db194fcd7b63",
"date_created": "2026-02-17T05:53:12.008000Z",
"date_modified": "2026-02-17T05:53:12.008000Z",
"file_hash": "07af489367c714839d6bf4086b6590a336cc889223022285b428d6ac8a95586b",
"private": false,
"record": {
"abstract": "The inherent safety alignment of Large Language Models (LLMs) is prone to erosion during fine-tuning, even when using seemingly innocuous datasets. While existing defenses attempt to mitigate this via data selection, they typically rely on heuristic, instance-level assessments that neglect the global geometry of the data distribution and fail to explicitly repel harmful patterns. To address this, we introduce Safety Optimal Transport (SOT), a novel framework that reframes safe fine-tuning from an instance-level filtering challenge to a distribution-level alignment task grounded in Optimal Transport (OT). At its core is a dual-reference ``push-pull\u0027\u0027 weight-learning mechanism: SOT optimizes sample importance by actively pulling the downstream distribution towards a trusted safe anchor while simultaneously pushing it away from a general harmful reference. This establishes a robust geometric safety boundary that effectively purifies the training data. Extensive experiments across diverse model families and domains demonstrate that SOT significantly enhances model safety while maintaining competitive downstream performance, achieving a superior safety-utility trade-off compared to baselines.",
"arxiv_id": "2601.07200",
"authors": [
"Haozhong Wang",
"Zhuo Li",
"Yibo Yang",
"He Zhao",
"Hongyuan Zha",
"Dandan Guo"
],
"categories": [
"cs.LG",
"cs.AI"
],
"license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/",
"title": "Safeguarding LLM Fine-tuning via Push-Pull Distributional Alignment",
"url": "https://arxiv.org/abs/2601.07200",
"version": "v1"
},
"schema_id": "dorsal/arxiv",
"source": {
"execution_id": "b7863894-e38b-4b3f-983e-c9c00a5077ba",
"id": "arXiv Dataset",
"type": "Model",
"variant": "snapshot-2026-01-17",
"version": "0.1.0"
},
"user_id": 1000002
}