🛰️ Daily AI Frontier
‹ back to 2026-09-10

Active Adaptation, Not Static Defense: Temporal Dynamics of Preventative Steering in Adversarial Fine-Tuning

Research LLMs & Foundation Models

Ranking

Overall 78
Content 95
Popularity 37

Observed public metrics from 1 member.

Merged summary

TL;DR - This paper finds that preventative steering protects LLMs from malicious fine-tuning through active, early-stage adaptation rather than a persistent static weight offset. It proposes Progressive Intensity Scheduling (PIS), which improves safety robustness while reducing harmful trait expression across evaluated Qwen2.5 and Gemma-3 models.

  • Defensive adaptation occurs early in fine-tuning, followed by a steady state in which the corrective signal decays.
  • Attention output projections are identified as the dominant residual-stream write route for defensive parameter updates.
  • Preserving or reinjecting the learned weight offset does not maintain protection, suggesting that static parameter changes are insufficient.
  • PIS raises steering intensity after fixed-strength alignment starts degrading, outperforming static-strength steering in the reported evaluations.

Sources (1)

Active Adaptation, Not Static Defense: Temporal Dynamics of Preventative Steering in Adversarial Fine-Tuning

arXiv cs.CL Jing Guan, Yachao Yang, Zhaoliang Liu, Yuyao Zhang, Fanyu Meng, Junlan Feng 2026-09-09 arXiv:2609.10142
Public signals Semantic Scholar citations 0 · Semantic Scholar influential citations 0
Providers: Hugging Face · N/A OpenAlex · N/A Publisher · N/A Semantic Scholar · Citations 0 · Influential citations 0 X · N/A Fetched 2026-09-18 14:16:44.436812 UTC

TL;DR - This paper finds that preventative steering protects LLMs from malicious fine-tuning through active, early-stage adaptation rather than a persistent static weight offset. It proposes Progressive Intensity Scheduling (PIS), which improves safety robustness while reducing harmful trait expression across evaluated Qwen2.5 and Gemma-3 models.

  • Defensive adaptation occurs early in fine-tuning, followed by a steady state in which the corrective signal decays.
  • Attention output projections are identified as the dominant residual-stream write route for defensive parameter updates.
  • Preserving or reinjecting the learned weight offset does not maintain protection, suggesting that static parameter changes are insufficient.
  • PIS raises steering intensity after fixed-strength alignment starts degrading, outperforming static-strength steering in the reported evaluations.
item →