Salvato in:
| Autori principali: | , , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2026
|
| Soggetti: | |
| Accesso online: | https://arxiv.org/abs/2601.19487 |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866911642234650624 |
|---|---|
| author | Zhang, Haonan Wang, Dongxia Liu, Yi Chen, Kexin Wang, Wenhai |
| author_facet | Zhang, Haonan Wang, Dongxia Liu, Yi Chen, Kexin Wang, Wenhai |
| contents | Safety-aligned LLMs suffer from two failure modes: jailbreak (answering harmful inputs) and over-refusal (declining benign queries). Existing vector steering methods adjust the magnitude of answer vectors, but this creates a fundamental trade-off -- reducing jailbreak increases over-refusal and vice versa. We identify the root cause: LLMs encode the decision to answer (answer vector $v_a$) and the judgment of input safety (benign vector $v_b$) as nearly orthogonal directions, treating them as independent processes. We propose LLM-VA, which aligns $v_a$ with $v_b$ through closed-form weight updates, making the model's willingness to answer causally dependent on its safety assessment -- without fine-tuning or architectural changes. Our method identifies vectors at each layer using SVMs, selects safety-relevant layers, and iteratively aligns vectors via minimum-norm weight modifications. Experiments on 12 LLMs demonstrate that LLM-VA achieves 11.45% higher F1 than the best baseline while preserving 95.92% utility, and automatically adapts to each model's safety bias without manual tuning. Code and models are available at https://hotbento.github.io/LLM-VA-Web/. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2601_19487 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | LLM-VA: Resolving the Jailbreak-Overrefusal Trade-off via Vector Alignment Zhang, Haonan Wang, Dongxia Liu, Yi Chen, Kexin Wang, Wenhai Machine Learning Artificial Intelligence Safety-aligned LLMs suffer from two failure modes: jailbreak (answering harmful inputs) and over-refusal (declining benign queries). Existing vector steering methods adjust the magnitude of answer vectors, but this creates a fundamental trade-off -- reducing jailbreak increases over-refusal and vice versa. We identify the root cause: LLMs encode the decision to answer (answer vector $v_a$) and the judgment of input safety (benign vector $v_b$) as nearly orthogonal directions, treating them as independent processes. We propose LLM-VA, which aligns $v_a$ with $v_b$ through closed-form weight updates, making the model's willingness to answer causally dependent on its safety assessment -- without fine-tuning or architectural changes. Our method identifies vectors at each layer using SVMs, selects safety-relevant layers, and iteratively aligns vectors via minimum-norm weight modifications. Experiments on 12 LLMs demonstrate that LLM-VA achieves 11.45% higher F1 than the best baseline while preserving 95.92% utility, and automatically adapts to each model's safety bias without manual tuning. Code and models are available at https://hotbento.github.io/LLM-VA-Web/. |
| title | LLM-VA: Resolving the Jailbreak-Overrefusal Trade-off via Vector Alignment |
| topic | Machine Learning Artificial Intelligence |
| url | https://arxiv.org/abs/2601.19487 |