Saved in:
Bibliographic Details
Main Authors: Lan, Wenhao, Li, Shan, Lai, Xinhua, Wu, Meiqi, Yang, Junbin, Shen, Haihua, Yang, Yijun
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2604.27019
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910256459677696
author Lan, Wenhao
Li, Shan
Lai, Xinhua
Wu, Meiqi
Yang, Junbin
Shen, Haihua
Yang, Yijun
author_facet Lan, Wenhao
Li, Shan
Lai, Xinhua
Wu, Meiqi
Yang, Junbin
Shen, Haihua
Yang, Yijun
contents Safety-aligned language models must refuse harmful requests without broad over-refusal, but it remains unclear how dynamic adversarial fine-tuning changes refusal-control carriers: Kullback--Leibler (KL)-constrained directions or small subspaces that causally modulate refusal without large safe-prompt distribution shifts. We study a 7B backbone under supervised fine-tuning (SFT) and Robust Refusal Dynamic Defense (R2D2), aligning HarmBench, StrongREJECT, and XSTest evaluations with five-anchor geometry measurements, causal interventions, and sparse adaptive stress tests. R2D2 drives fixed-source HarmBench attack success to zero at early checkpoints; however, these checkpoints also exhibit maximal XSTest refusal and fail a benign-utility audit. Later checkpoints partially recover utility-facing behavior while reopening attack success, with adaptive GCG attack success rate rising to 0.415 at step 250 and 0.613 at step 500. Internally, R2D2 preserves a late-layer admissible refusal-control carrier through step 100 and then relocates the best admissible carrier to an early layer; SFT relocates earlier yet remains less robust. Effective rank stays near 1.24, and SFT shows larger principal-angle drift, arguing against both dimensional expansion and drift magnitude as sufficient explanations. Causal interventions support a low-dimensional but utility-coupled carrier. These results support a geometry-reorganization account of R2D2 along a robustness--utility frontier, without establishing adaptive robustness.
format Preprint
id arxiv_https___arxiv_org_abs_2604_27019
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Dynamic Adversarial Fine-Tuning Reorganizes Refusal Geometry
Lan, Wenhao
Li, Shan
Lai, Xinhua
Wu, Meiqi
Yang, Junbin
Shen, Haihua
Yang, Yijun
Machine Learning
Computation and Language
Cryptography and Security
Safety-aligned language models must refuse harmful requests without broad over-refusal, but it remains unclear how dynamic adversarial fine-tuning changes refusal-control carriers: Kullback--Leibler (KL)-constrained directions or small subspaces that causally modulate refusal without large safe-prompt distribution shifts. We study a 7B backbone under supervised fine-tuning (SFT) and Robust Refusal Dynamic Defense (R2D2), aligning HarmBench, StrongREJECT, and XSTest evaluations with five-anchor geometry measurements, causal interventions, and sparse adaptive stress tests. R2D2 drives fixed-source HarmBench attack success to zero at early checkpoints; however, these checkpoints also exhibit maximal XSTest refusal and fail a benign-utility audit. Later checkpoints partially recover utility-facing behavior while reopening attack success, with adaptive GCG attack success rate rising to 0.415 at step 250 and 0.613 at step 500. Internally, R2D2 preserves a late-layer admissible refusal-control carrier through step 100 and then relocates the best admissible carrier to an early layer; SFT relocates earlier yet remains less robust. Effective rank stays near 1.24, and SFT shows larger principal-angle drift, arguing against both dimensional expansion and drift magnitude as sufficient explanations. Causal interventions support a low-dimensional but utility-coupled carrier. These results support a geometry-reorganization account of R2D2 along a robustness--utility frontier, without establishing adaptive robustness.
title Dynamic Adversarial Fine-Tuning Reorganizes Refusal Geometry
topic Machine Learning
Computation and Language
Cryptography and Security
url https://arxiv.org/abs/2604.27019