Representation Bending for Large Language Model Safety
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2025
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866912482306555904 |
|---|---|
| author | Yousefpour, Ashkan Kim, Taeheon Kwon, Ryan S. Lee, Seungbeen Jeung, Wonje Han, Seungju Wan, Alvin Ngan, Harrison Yu, Youngjae Choi, Jonghyun |
| author_facet | Yousefpour, Ashkan Kim, Taeheon Kwon, Ryan S. Lee, Seungbeen Jeung, Wonje Han, Seungju Wan, Alvin Ngan, Harrison Yu, Youngjae Choi, Jonghyun |
| contents | Large Language Models (LLMs) have emerged as powerful tools, but their inherent safety risks - ranging from harmful content generation to broader societal harms - pose significant challenges. These risks can be amplified by the recent adversarial attacks, fine-tuning vulnerabilities, and the increasing deployment of LLMs in high-stakes environments. Existing safety-enhancing techniques, such as fine-tuning with human feedback or adversarial training, are still vulnerable as they address specific threats and often fail to generalize across unseen attacks, or require manual system-level defenses. This paper introduces RepBend, a novel approach that fundamentally disrupts the representations underlying harmful behaviors in LLMs, offering a scalable solution to enhance (potentially inherent) safety. RepBend brings the idea of activation steering - simple vector arithmetic for steering model's behavior during inference - to loss-based fine-tuning. Through extensive evaluation, RepBend achieves state-of-the-art performance, outperforming prior methods such as Circuit Breaker, RMU, and NPO, with up to 95% reduction in attack success rates across diverse jailbreak benchmarks, all with negligible reduction in model usability and general capabilities. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2504_01550 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Representation Bending for Large Language Model Safety Yousefpour, Ashkan Kim, Taeheon Kwon, Ryan S. Lee, Seungbeen Jeung, Wonje Han, Seungju Wan, Alvin Ngan, Harrison Yu, Youngjae Choi, Jonghyun Machine Learning Computation and Language Cryptography and Security Large Language Models (LLMs) have emerged as powerful tools, but their inherent safety risks - ranging from harmful content generation to broader societal harms - pose significant challenges. These risks can be amplified by the recent adversarial attacks, fine-tuning vulnerabilities, and the increasing deployment of LLMs in high-stakes environments. Existing safety-enhancing techniques, such as fine-tuning with human feedback or adversarial training, are still vulnerable as they address specific threats and often fail to generalize across unseen attacks, or require manual system-level defenses. This paper introduces RepBend, a novel approach that fundamentally disrupts the representations underlying harmful behaviors in LLMs, offering a scalable solution to enhance (potentially inherent) safety. RepBend brings the idea of activation steering - simple vector arithmetic for steering model's behavior during inference - to loss-based fine-tuning. Through extensive evaluation, RepBend achieves state-of-the-art performance, outperforming prior methods such as Circuit Breaker, RMU, and NPO, with up to 95% reduction in attack success rates across diverse jailbreak benchmarks, all with negligible reduction in model usability and general capabilities. |
| title | Representation Bending for Large Language Model Safety |
| topic | Machine Learning Computation and Language Cryptography and Security |
| url | https://arxiv.org/abs/2504.01550 |